I am hardly a kernel programmer, but it seems the real issue here is still that the release_sock() call isn't actually doing what he assumed.
While I was making a baseless assumption, I like to believe that it was baseless because there was not enough available information, and not because I was ignoring obvious sources of information :P
[1]: http://lxr.free-electrons.com/ident?v=4.1;i=release_sock [2]: https://github.com/multipath-tcp/mptcp/blob/e77e2ca0b5339af2...